Black box auditing of language models is crucial for pre-deployment, yet it may overlook subtle misalignments and hidden information.
To better elicit hidden information during the auditing process, we introduce \emph{overthinking}: the process of using reasoning task vectors to amplify the propensity of reasoning models to think out loud.
Given the parameters of a non-reasoning instruct model $M$ and a reasoning-distilled model @@@MATH_BLOCK1@@@, we define the \emph{overthinking model} as: $$ \boldsymbol{\theta}{\mathcal{O}\alpha} = \boldsymbol{\theta}{\mathcal{M}} + \alpha(\boldsymbol{\theta}{\mathcal{R}} - \boldsymbol{\theta}{\mathcal{M}}) $$ where $\alpha > 1$ amplifies reasoning beyond the pure reasoning model $R$.
Additionally, we introduce new layer-wise attenuation strategies that selectively amplify reasoning without losing the quality and coherence of model outputs.
We demonstrate that overthinking models are more likely to reveal hidden information across four experimental settings, suitable for 2B-32B models.
Our findings suggest that reasoning amplification may surface secrets or unintended behaviors acquired during training up to 10 times more frequently than the original reasoning model.
How secrets surface depends on the secret type: some require perturbation along the reasoning direction, while others yield to any sufficiently large weight perturbation.
Blogger's Review: The overthinking approach significantly enhances the ability to reveal hidden information by amplifying reasoning weights. This innovation provides a powerful tool for auditing language models, especially valuable in understanding model behaviors and potential issues.